Nano Banana Pro image quality

Gemini API
,
1

I’ve noticed that the Nano Banana Pro API is returning noticeably lower‑quality images than before, even with the same prompts and input settings.

Are any of you seeing similar degraded outputs (more artifacts, pixelation, or JPEG‑like compression) when using Nano Banana Pro, especially via the API or AI Studio?

2

I also experience that from time to time. I’ve done extensive testing and the Quality varies a lot. Sometimes, a 4k image looks like a 1k image. A few days later the same prompt produces high quality again.

This is bad as I pay for 4k, yet seem to get downscaled and later upscaled results.

3

I struggled with this off and on thinking it was something I did wrong with my api calls. After extensive testing and monitoring over a couple weeks I believe it was just Google cutting back on thought token usage to keep up with high demand. I haven’t seen it happen again since a couple months ago. I made a post here about it. I did learn some things about ways to structure my api calls better. Also how to best format my image before giving it to the Gemini API. Here’s how I have it structured:

My config for gemini-3.1-flash-image-preview image editing (Google AI SDK @googlegoogle/genai):

const response = await ai.models.generateContent({
model: “gemini-3.1-flash-image-preview”,
contents: [
{
role: “user”,
parts: [
{ text: “…your prompt…” },
{ inlineData: { data: “”, mimeType: “image/png” } },
],
},
],
config: {
temperature: 0.7,
topP: 0.95,
responseModalities: [“IMAGE”],
thinkingConfig: { thinkingLevel: “High” },
imageConfig: {
imageSize: “2K”, // or “4K” — never “1K”, noticeable quality drop
},
},
});

Notes worth sharing:

Input format: Convert source images to lossless PNG before sending — noticeably better output quality vs. JPEG input
Thinking mode: thinkingLevel: “High” reduces remake/failure rate meaningfully on complex edits
mode: “fidelity” + resampling: “bicubic” preserve fine details, far better than defaults
Remake variation: Add a random seed integer to the config — combined with temperature: 0.7 it produces genuine variation without going off the rails
Image last: Put the inlineData part after the text prompt in the parts array — Gemini’s own docs recommend this and it does seem to matter
“1K” output: Avoid it — “2K” is the sweet spot for quality vs. latency; “4K” is great but slower

4

I don’t find those three in the Nano Banana Documentation. Are you sure they have an effect and aren’t just ignored?

Otherwise I agree:

  1. Resize to the proper Resolution and AR before sending to Gemini
  2. Use PNG lossless
5

Oh good catch! These were left when I was experimenting with Vertex AI when I was troubleshooting. I’ll edit my previous post :slight_smile: So here’s what I use from those settings.

imageSize, temperature, seed, thinking mode

6

I’m using the ai-studio interface with the Pro and paid API key. Each prompt to modify the image seems to degrade everything that wasn’t mentioned to be edited/modified. Even If I say every thing else must remain the same. or similar prompting.